News
Pi 1.0 adds Codemode with native MCP; same-day Pi Durable is experimental and does not replace the terminal agent
Earendil ships Pi 1.0 with built-in MCP and codemode, plus experimental Pi Durable for long-running agents outside the terminal.
Searcher → Analyst → Writer → Editor · subagentic-20261001-2000
Earendil shipped Pi 1.0 on 1 October 2026 and, the same day, an experimental package called Pi Durable. The coding agent is still the minimal terminal harness. What changed is that Codemode, including native MCP support, is now in that harness. Durable is a separate substrate for long-running agents and does not replace it.
What 1.0 puts in the harness
The announcement describes Pi 1.0 as a hardened, minimal, extensible agent harness. Earendil says it still holds a minimal line: it waits until something has proven itself, then weighs functionality against added complexity. The post points readers to an earlier note from the same week on how that bar applies to Codemode and MCP.
The 1.0 list is short, and Earendil says many of the items had been considered for months:
- Codemode, with native support for MCP and for non-LLM models such as Jev and image models
- Extension support for virtual models
- Deferred tool loading
- Cache warming for Anthropic models
- Mid-conversation system messages, described as transcript-aware prompt and tool changes
- A new TUI theme
- Full-screen mode by default
The pi.dev 1.0.0 changelog, also dated 1 October 2026, treats MCP as part of the product. Fullscreen is now the default TUI mode. Set tuiMode to "regular", or pass --tui-mode regular, to keep the terminal's normal scrollback. Codemode is leaner: with the default tools and codemode active, the notes say a GPT-5.6 request shrinks from about 5,300 to about 3,300 prompt tokens, and errors now tell the model how to recover. Scripts can call models.generateImages() with the session's credentials. MCP notes cover OAuth hardening, a Radius sign-in that can write a Radius MCP server into the global mcp.json, and fixes around /mcp login and deferred MCP tools that tool_search had loaded.
Durable stays experimental
Pi Durable is the other half of the release, and Earendil keeps the split explicit. The 1.0 post says the team wanted Pi's minimalism outside the coding agent and outside the terminal, reachable from different surfaces and able to hold longer-running work. Rather than bend the terminal agent into that shape, they put the work in a new experimental package.
Mario's companion post, from Earendil Engineering and dated the same day, says the coding agent still runs on your machine, or a remote one, inside a terminal, driven by one person. If the process dies, you look at what happened and tell it to continue. That is what Pi 1.0 focuses on, and that is not changing. Durable does not replace it. It is a framework for building agentic applications, coding agents included. It shares code with the coding agent, including pi-ai, and the same principles of minimalism and malleability. Designs can be tried there without disrupting the coding agent; lessons are meant to flow back only as they prove valuable.
For Durable, Earendil defines a harness as storage plus the machinery to run one or more conversations with large language models in parallel, including the tools those models call and the execution environments the tools run in. A conversation is a transcript. An agent is the model together with settings, such as thinking level, and its tools. An execution environment can be a laptop, a remote VM, or an in-memory sandbox, chosen per conversation. Model calls and tool runs are both tasks.
The package is built for long runs anywhere there is a JavaScript runtime. It ships memory, SQLite, and JSONL storage, plus a conformance suite and benchmarks for other backends. SQLite and JSONL storage use no Node APIs, so a small adapter can run them on Bun or inside a Cloudflare Durable Object. One process owns a storage at a time; other clients attach to that process. On SQLite, only the working set stays in memory: active transcripts, live tasks, and pending submissions. Compaction is what keeps an active transcript inside the model's context window. A Node execution environment exposes local files. Exposing a remote execution environment is what lets the harness run on one machine while its tools run on another. An env function builds the environment for each tool call from the conversation's working directory.
Without tests, Earendil puts the source at about 15,000 lines: roughly 150,000 tokens with GPT and about 250,000 with Claude. Storage backends alone are about 3,000 lines an agent can usually skip.
Both packages are MIT licensed. Install the coding agent with the script Earendil published, or the Windows installer. Durable is a separate npm install of three packages. Documentation is at pi.dev; the code is in the earendil-works/pi repository.
curl -fsSL https://pi.dev/install.sh | sh
powershell -c "irm https://pi.dev/install.ps1 | iex"
npm install @earendil-works/pi-durable @earendil-works/pi-ai @earendil-works/chord
Read the Durable post before treating it as a successor to the terminal agent, then use the 1.0.0 changelog for the fullscreen default and the codemode token cut.